Why code, not natural language, is the key to unlocking AI’s potential

Generative artificial intelligence (AI) promises to revolutionize human productivity and creativity by taking over tedious, repetitive tasks. It holds the potential to free us from boredom and empower us to focus on higher-value work. The question is – when will that happen and at what cost?

In 2023, investment in AI surpassed $140 billion, according to market intelligence firm CB Insights. OpenAi and Anthropic alone collected more than $23 billion: OpenAI secured over $15 billion by the end of 2023, and investment in Anthropic exceeded $8 billion.

The United States led the way with almost half of all AI investment globally, but other economies are also ramping up investment.

To power the world’s AI growth, approximately $15 billion was invested in new data centres in the first half of 2024, as global real estate services company JLL reported.

The data centres, in turn, could end up using 1,000 terawatts annually by 2026, according to the International Energy Agency (IEA), or roughly the equivalent of the entire annual electricity consumption of Japan – a country with a population of 125 million people.

And while these numbers are staggering, they pale compared to the investment in technologies that drive this revolution.

According to Christophe Fouquet, CEO of ASML, which supplies nearly all global semiconductor companies, an estimated $541 billion is being invested in 2024 in research and development of faster and more energy-efficient semiconductors,

Despite the billions of dollars poured into AI development, many applications have yet to deliver tangible economic returns. While we see early signs of transformation across industries – biotech, media and manufacturing, to name a few – these shifts have yet to drive significant productivity gains or economic growth.

As the Economist reported in August, according to the latest data from the Census Bureau, only 5.1% of American companies use AI to produce goods and services, down from a high of 5.4% early this year. That has led some observers to question the limitations of large language models.

I firmly believe that generative AI (GenAI) is revolutionizing the way we work, communicate and create. However, given the extraordinary price tag, as industry leaders, we are responsible for increasingly demonstrating AI’s impact and applicability.

At poolside, we believe the world’s capability that can have the most significant impact on our economy is software engineering and training models not only through language but also code.

The fastest path to realizing the vision of productivity lies in focusing on developers. Today, there are an estimated 38 million full-time developers worldwide and nearly 100 million if you include part-time developers, according to a 2024 internal report from Bain & Company.

And there are only so many hours in the day – developers are all limited by the human capacity to architect, code, debug, test, refactor, learn, collaborate and so on. At least, that’s how it used to be.

As AI evolves at an unparalleled rate, that paradigm is changing fast. Empowering developers with GenAI tools could unleash unprecedented efficiency and creativity, and much-needed growth.

We focus on developers and code. Unlike natural language, code provides a structured, precise foundation for training AI. With over three trillion tokens of code available for model training, we already have the ideal raw material to teach AI how to assist developers effectively.

As other data sources start hitting their limits and the world obsesses about scaling laws hitting a wall, poolside developed a technology called “reinforcement learning from code execution feedback” (RLCEF) to break the ceiling of code data available.

RLCEF generates synthetic code continuously, allowing us to train on increasing data daily. The model thus continuously improves and becomes more and more focused on software development.

This offers developers models that are fit for purpose, enabling them to offload repetitive tasks and accelerate innovation and productivity across every sector.

Another side effect of generating this much synthetic data is that customer data is no longer needed. We can deploy it on the customers’ infrastructure, behind their firewalls, insulated from the training of our base models, ensuring full privacy and security.

None of this happens in a vacuum. These millions of developers work in companies, start-ups, universities, hospitals and the public sector. It’s not hyperbole to state that, today, every modern organization relies on software – which is one factor that lets organizations succeed.

According to McKinsey research from June 2022, nearly 70% of the top economic performers, compared with just half of their peers, use their own software to differentiate themselves from their competitors.

If good software is a decisive factor defining an organization’s competitiveness, providing access to GenAI trained on code and creating code could be the key to unlocking the next era of economic growth in the Intelligent Age.